Why Rakudo needs NQP

Rakudo, a popular Perl 6 compiler, is built on top of a smaller compiler called "NQP", short for Not Quite Perl.

Reading through a recent ramble by chromatic, I felt like he said "Rakudo needs NQP to be able to ditch Parrot, once NQP runs on a different platform" (NQP is the "another layer", which sits between Rakudo and Parrot, mentioned in the next-to-final paragraph).

I'm sure chromatic knows that VM independence is the least important reason for having NQP at all, but the casual reader might not, so let me explain the real importance of NQP for Rakudo here.

The short version is just a single word: bootstrapping.

The longer version is that large parts of Rakudo are written in Perl 6 itself (or a subset thereof), and something is needed to break the circularity.

In particular the base of the compiler is written in a subset of Perl 6, and NQP compiles those parts to bytecode, which can then compile the rest of the compiler.

This is not just because we have a fancy for Perl 6, and thus want to write as much of the code in Perl 6, but there are solid technical reasons for writing the compiler in Perl 6.

In Perl 6, the boundary between run time and compile time is blurred, as well as the boundary between the compiler, the run time library and user-space code. For example you alter the grammar with which your source code is parsed, by injecting your own grammar rules.

"Your own grammar rules" above refers to user-space code, while the grammar that is being altered is part of the compiler. If we had written the compiler in something else than Perl 6 (for example Java), it would be horribly difficult to inject user-space Perl 6 code into compiled code from a different language.

And the code not only needs to be injected, but the data passed back and forth between the compiler and the user space need to be Perl 6 objects, so all important data structures in the compiler need to be Perl 6 based anyway.

And it's not just for grammar modifications: At its heart, Perl 6 is an object oriented language. When the compiler sees a class definition, it translates them to a series of method calls on the metaobject, which again needs to be a Perl 6 object, otherwise it wouldn't be easily usable and extensible from the user space.

Now you might think that grammar modifications and changes to the Metaobject are pretty obscure features, and you could get along just fine with an incomplete Perl 6 compiler that neglected those two areas. But even then you'd have lots of interactions between run time and compile time. For example consider a numeric literal like 42. Obviously that needs to be constructed of type Int. What's less obvious is that it needs to be constructed to be of type Int at compile time already, because Perl 6 code can run interleaved with the compilation. So the compiler needs to be able to handle Perl 6 objects in all their generality, which is a huge pain if the compiler is not written in Perl 6.

Rakudo has cheated on that front in the past, and consequently has had lots of bugs and limitations due to non-Perl 6 objects leaking out at unexpected ends. If you ever got a "Null PMC Access" from Rakudo, you know what I mean.

The lesson we learned was that you need a Perl 6 compiler to implement a Perl 6 compiler, even if that first Perl 6 compiler can handle only a rather limited subset of Perl 6.

And there are also quite some benefits to this approach. For example NQP's new regex engine is implemented as a role in NQP. It is mixed into an NQP class which allows us to build Rakudo, but it is also mixed in a Perl 6 class, which allows the generation of Perl 6-level Match objects without any need to create NQP-level match objects first, and then wrap them in Perl 6 Match objects.

That's what NQP does for us. It allows us to actually write a Perl 6 compiler.